Vous êtes sur la page 1sur 20

Chapter 7

Theoretical Foundations
of the Finite Element
Method

Using finite element methods, we need to answer these questions:


• What is the appropriate functional space V for the solution?
• What is the appropriate weak or variational form of a differential equation?
• What kind of basis functions or finite element spaces should we choose?
• How accurate is the finite element solution?
We briefly address these questions in this Chapter. Recalling that finite element methods
are based on integral forms and not on the pointwise sense as in finite difference methods.
We will generalize the theory corresponding to the pointwise form to deal with integral
forms.

7.1 Functional spaces


A functional Space is a set of functions with operations. For example,
n o
C(Ω) = C 0 (Ω) = u(x), u(x) is continuous on Ω (7.1)

is a linear space that contains all continuous functions on Ω, the domain where the functions
are defined, i.e., Ω = [0, 1]. The space is linear because for any real numbers α and β and
u1 ∈ C(Ω) and u2 ∈ C(Ω), we have αu1 + βu2 ∈ C(Ω).
The functional space with first order continuous derivatives in 1D is

C 1 (Ω) = u(x), u(x), u′ (x) are continuous on Ω , (7.2)

and similarly

C m (Ω) = u(x), u(x), u′ (x), · · · , u(m) are continuous on Ω . (7.3)

Obviously,

C0 ⊃ C1 ⊃ · · · ⊃ Cm ⊃ · · · . (7.4)

155

i i

i i
156 Chapter 7. Theoretical Foundations of the Finite Element Method

Then as m → ∞, we define

C ∞ (Ω) = {u(x), u(x) is indefinitely differentiable on Ω} . (7.5)

For example, ex , sin x, cos x, and polynomials, are in C ∞ (−∞, ∞), but some other ele-
mentary functions such as log x, tan x, cot x are not if x = 0 is in the domain.

7.1.1 Multi-dimensional spaces and multi-index notations


Let us now consider multi-dimensional functions u(x) = u(x1 , x2 , · · · , xn ), x ∈ Rn , and a
corresponding multi-index notation that simplifies expressions for partial derivatives. We
can write α = (α1 , α2 , · · · , αn ), αi ≥ 0 for an integer vector in Rn , e.g., if n = 5, then
α = (1, 2, 0, 0, 2) is one of possible vectors. We can readily represent a partial derivative
as

∂ |α| u
Dα u(x) = , |α| = α1 + α2 + · · · + αn , αi ≥ 0 (7.6)
∂xα α2
1 ∂x2
1
· · · ∂xα
n
n

which is the so called multi-index notation.

Example 7.1. For n = 2 and u(x) = u(x1 , x2 ), all possible Dα u when |α| = 2 are

∂2u
α = (2, 0), Dα u = ,
∂x21
∂2u
α = (1, 1), Dα u = ,
∂x1 ∂x2
∂2u
α = (0, 2), Dα u = .
∂x22

With the multi-index notation, the C m space in a domain Ω ∈ Rn can be defined as

C m (Ω) = {u(x1 , x2 , · · · , xn ), Dα u are continuous on Ω, |α| ≤ m} (7.7)

i.e., all possible derivatives up to order m are continuous on Ω.

Example 7.2. For n = 2 and m = 3, we have u, ux , uy , uxx , uxy , uyy , uxxx , uxxy , uxyy
and uyyy all continuous on Ω if u ∈ C 3 (Ω), or simply Dα u ∈ C 3 (Ω) for |α| ≤ 3. Note
that C m (Ω) has infinite dimensions.

The distance in C 0 (Ω) is defined as

d(u, v) = max |u(x) − v(x)|


x∈Ω

with the properties (1), d(u, v) ≥ 0 ; (2), d(u, v) = 0 if and only if u ≡ v ; and (3),
d(u, v + w) ≤ d(u, v) + d(u, w), the triangle inequality. A linear space with a distance
defined is called a metric space.

i i

i i
7.2. Spaces for integral forms, L2 (Ω) and Lp (Ω) 157

A norm in C 0 (Ω) is a non-negative function of u that satisfies

ku(x)k = d(u, θ) = max |u(x)| , where θ is the zero element,


x∈Ω

with the properties (1), ku(x)k ≥ 0 , and ku(x)k = 0 if and only if u ≡ 0 ;


(2), kαu(x)k = |α|ku(x)k, where α is a number ;
(3), ku(x) + v(x)k ≤ ku(x)k + kv(x)k , the triangle inequality.

A linear space with a norm defined is called a normed space. In C m (Ω), the distance and
the norm are defined as

d(u, v) = max max |Dα u(x) − Dα v(x)| , (7.8)


0≤|α|≤m x∈Ω

ku(x)k = max max |Dα u| . (7.9)


0≤|α|≤m x∈Ω

7.2 Spaces for integral forms, L2 (Ω) and Lp (Ω)


In analogy to pointwise spaces C m (Ω), we can define Sobolev spaces H m (Ω) in integral
forms. The square-integrable space H 0 (Ω) = L2 (Ω) is defined as
 Z 
L2 (Ω) = u(x), u2 (x) dx < ∞ (7.10)

corresponding to the pointwise C 0 (Ω) space. It is easy to see that

C(0, 1) = C 0 (0, 1) ⊂ L2 (0, 1) .

Example 7.3. It is easy to verify that y(x) = 1/x1/4 ∈


/ C 0 (0, 1), but
Z 1  2 Z 1
1 1
dx = √ dx = 2 < ∞
0
x1/4 0 x

so that y(x) ∈ L2 (0, 1). But it is obvious that y(x) is not in C(0, 1) since y(x) blows up as
x → 0 from x > 0, see Fig.7.1 in which we also show that a piecewise constant function 0
and 1 in (0, 1) is in L2 (0, 1) but not in C(0, 1) since the function is discontinuous (non-
removable discontinuity) at x = 0.5.

The distance in L2 (Ω) is defined as


Z 1/2
2
d(f, g) = |f − g| dx , (7.11)

which satisfies the three conditions of the distance definition; so L2 (Ω) is a metric space.
We say that two functions f and g are identical (f ≡ g) in L2 (Ω) if d(f, g) = 0. For
example, the following two functions are identical in L2 (−2, 2):
( (
0, if −2 ≤ x < 0, 0, if −2 ≤ x ≤ 0,
f (x) = g(x) =
1, if 0 ≤ x ≤ 2, 1, if 0 < x ≤ 2.

i i

i i
158 Chapter 7. Theoretical Foundations of the Finite Element Method

2.5

2
y=1/x1/4

1.5

1
y=1

0.5

y=0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 7.1. Plot of two functions that are in L2 (0, 1) but not in C[0, 1].
1
One function is y(x) = √4
x
. The other one is y(x) = 0 in [0, 1/2) and y(x) = 1 in
[1/2, 1].

The norm in the L2 (Ω) space is defined as


Z 1/2
2
kukL2 = kuk0 = |u| dx . (7.12)

It is straightforward to prove that the usual properties for the distance and the norm hold.
We say that L2 (Ω) is a complete space, meaning that any Cauchy sequence {fn (Ω)}
in L has a limit in L2 (Ω); i.e., there is a function f ∈ L2 (Ω) such that
2

lim kfn − f kL2 = 0, or lim fn = f.


n→∞ n→∞

A Cauchy sequence is a sequence that satisfies the property that, for any given positive
number ǫ, no matter how small it is, there is an integer N such that

kfn − fm kL2 < ǫ, if m ≥ N, n≥N.

A complete normed space is called a Banach space (a Cauchy sequence converges in terms
of the norm), so L2 (Ω) is a Banach space.

7.2.1 The inner product in L2


For any two vectors
   
x1 y1
   
 x2   y2 
   
x=
 ..  ,
 y=


.. 
 .   . 
   
xn yn

i i

i i
7.2. Spaces for integral forms, L2 (Ω) and Lp (Ω) 159

in Rn , we recall that the inner product is


X
n

(x, y) = xT y = x i yi = x 1 y1 + x 2 y2 + · · · + x n yn .
i=1

Similarly, the inner product in L2 (Ω) space in R, the real number space, is defined as
Z
(f, g) = f (x)g(x) dx , for any f and g ∈ L2 (Ω) , (7.13)

and satisfies the familiar properties
(f, g) = (g, f ) ,
(αf, g) = (f, αg) = α(f, g), ∀α ∈ R,
(f, g + w) = (f, g) + (f, w)
for any f, g and w ∈ L (Ω). The norm, distance, and inner product in L2 (Ω) are related
2

as follows:
Z 1/2
p 2
kuk0 = kukL2 (Ω) = (u, u) = d(u, θ) = |u| dx . (7.14)

With the L2 (0, 1) inner product, for the simple model problem
−u′′ = f , 0 < x < 1, u(0) = u(1) = 0,
we can rewrite the weak form as
(u′ , v ′ ) = (f, v) , ∀v ∈ H01 (0, 1) ,
and the minimization form as
1 ′ ′
min F (v) : F (v) = (v , v ) − (f, v) .
v∈H01 (0,1) 2

7.2.2 The Cauchy-Schwartz inequality in L2 (Ω)


p
For a Hilbert space with the norm kuk = (u, u), the Cauchy-Schwartz inequality is
|(u, v)| ≤ kuk0 kvk0 . (7.15)
Examples of the Cauchy-Schwartz inequality corresponding to inner products in Rn and
in L2 spaces are:
n ( n )1/2 ( n )1/2
X X 2 X 2

x i yi ≤ xi yi ,
i=1 i=1 i=1

( n )1/2
X n √ X 2

xi ≤ n xi ,
i=1 i=1

Z Z 1/2 Z 1/2

f gdx ≤ f 2
dx g 2
dx ,

Ω Ω Ω
Z Z 1/2

f dx ≤ 2
f dx V ,

Ω Ω

i i

i i
160 Chapter 7. Theoretical Foundations of the Finite Element Method

where V is the volume of Ω.

A proof of the Cauchy-Schwartz inequality


Noting that (u, u) = kuk2 , we construct a quadratic function of α given u and v:

f (α) = (u + αv, u + αv) = (u, u) + 2α(u, v) + α2 (v, v) ≥ 0 .

The quadratic function is non-negative; hence the discriminant of the quadratic form
satisfies

∆ = b2 − 4ac ≤ 0, i.e., 4(u, v)2 − 4(u, u)(v, v) ≤ 0 or (u, v)2 ≤ (u, u)(v, v) ,

yielding the Cauchy-Schwartz inequality |(u, v)| ≤ kukkvk, on taking the square root of
both sides.
A complete Banach space with an inner product defined is called a Hilbert space.
Hence L2 (Ω) is a Hilbert space (linear space, inner product, complete).

Relationships between the spaces


The relationships (and relevant additional properties) in the hierarchy of defined spaces
may be summarized diagrammatically:
Metric Space (distance) =⇒ Normed Space (norm) =⇒ Banach space (complete)
=⇒ Hilbert space (inner product).

7.2.3 Lp (Ω) spaces


An Lp (Ω) space is defined as
 Z 
Lp (Ω) = u(x), |u(x)|p dx < ∞ , (7.16)

and the distance in Lp (Ω) is defined as


Z 1/p
d(f, g) = |f − g|p dx . (7.17)

An Lp (Ω) space has a distance and is complete, so it is a Banach space. However, it is not
a Hilbert space, because no corresponding inner product is defined unless p = 2.

7.3 Sobolev spaces and weak derivatives


Similar to C m (Ω) spaces, we use Sobolev spaces H m (Ω) to define function spaces with
derivatives involving integral forms. If there is no derivative, then the relevant Sobolev
space is
 Z 
0 2 2
H (Ω) = L (Ω) = v(x), |v| dx < ∞ . (7.18)

i i

i i
7.3. Sobolev spaces and weak derivatives 161

7.3.1 Definition of weak derivatives


If u(x) ∈ C 1 [0, 1], then for any function φ ∈ C 1 (0, 1) such that φ(0) = φ(1) = 0 we recall
Z 1 1 Z 1 Z 1
′ ′
u (x)φ(x) = uφ − u(x)φ (x) dx = − u(x)φ′ (x) dx , (7.19)
0 0 0 0

where φ(x) is a test function in C01 (0, 1). The first order weak derivative of u(x) ∈ L2 (Ω) =
H 0 (Ω) is defined to be a function v(x) satisfying
Z Z
v(x)φ(x)dx = − u(x)φ′ (x)dx (7.20)
Ω Ω

for all φ(x) ∈ C01 (Ω) with φ(0) = φ(1) = 0. If such a function exists, then we write
v(x) = u′ (x).

Example 7.4.

Consider the following function u(x)


 x 1

 2, if 0 ≤ x < ,
2
u(x) =

 1 − x, 1
if ≤ x ≤ 1,
2 2
It is obvious that u(x) ∈ C[0, 1]) but u′ (x) 6∈ C(0, 1) since the classic derivative does
not exist at x = 21 . Let φ(x) ∈ C 1 (0, 1) be any function that vanishes at two ends, i.e.,
φ(0) = φ(1) = 0, and has first order continuous derivtive on (0, 1). We carry out the
following integration by parts.
Z 1 Z 1 Z 1
2
φ′ udx = φ′ udx + φ′ udx
0 0 1
2
1 1 Z 1 Z 1
2 2

= φ(x) u(x) + φ(x) u(x) − φu dx + φu′ dx
01 1
2 0 2
  Z 1 Z 1
2 φ(x) −φ(x)
= φ(1/2) u(1/2−) − u(1/2+) − dx − dx
0
2 1 2
2
Z 1
=− ψ(x)φ(x)dx,
0

where we have used the property that φ(0) = φ(1) = 0 and ψ(x) is defined as
 1 1

 2, if 0 < x < ,
2
ψ(x) =

 −1, if 1
< x < 1,
2 2

which is what we would expect. In other words, we define the weak derivative of u(x) as
u′ (x) = ψ(x) which is an H 1 (0, 1) but not a C(0, 1) function.

i i

i i
162 Chapter 7. Theoretical Foundations of the Finite Element Method

Similarly, the m-th order weak derivative of u(x) ∈ H 0 (Ω) is defined as a function
v(x) satisfying
Z Z
v(x)φ(x)dx = (−1)m u(x)φ(m) (x)dx (7.21)
Ω Ω

for all φ(x) ∈ C0m (Ω) with φ(x) = φ′ (x) = · · · = φ(m−1) (x) = 0 for all x ∈ ∂Ω. If such a
function exists, then we write v(x) = u(m) (x).

7.3.2 Definition of Sobolev spaces H m (Ω)


The Sobolev space H 1 (Ω) defined as

H 1 (Ω) = v(x) , Dα v ∈ L2 (Ω) , |α| ≤ 1 (7.22)

involves first order derivatives, e.g.,


 Z b Z b 
1 2 ′ 2
H (a, b) = v(x) , a < x < b , v dx < ∞ , (v ) dx < ∞ ,
a a

and in two space dimensions, H 1 (Ω) is defined as


 
∂v ∂v
H 1 (Ω) = v(x, y) , v ∈ L2 (Ω) , ∈ L2 (Ω) , ∈ L2 (Ω) .
∂x ∂y

The extension is immediate to the Sobolev space of general dimension



H m (Ω) = v(x) , Dα v ∈ L2 (Ω) , |α| ≤ m . (7.23)

7.3.3 Inner products in H m (Ω) spaces


The inner product in H 0 (Ω) is the same as that in L2 (Ω), i.e.,
Z
(u, v)H 0 (Ω) = (u, v)0 = uv dx .

The inner product in H 1 (a, b) is defined as


Z b

(u, v)H 1 (a,b) = (u, v)1 = uv + u′ v ′ dx ;
a

the inner product in H 1 (Ω) of two variables is defined as


ZZ  
∂u ∂v ∂u ∂v
(u, v)H 1 (Ω) = (u, v)1 = uv + + dxdy ; and

∂x ∂x ∂y ∂y

the inner product in H m (Ω) (of general dimension) is


ZZ X
(u, v)H m (Ω) = (u, v)m = (Dα u(x)) (Dα v(x)) dx . (7.24)
Ω |α|≤m

i i

i i
7.3. Sobolev spaces and weak derivatives 163

The norm in H m (Ω) (of general dimension) is


 1/2
Z X 
kukH m (Ω) = kukm = |Dα u(x)|2 dx , (7.25)
 Ω |α|≤m 

therefore H m (Ω) is a Hilbert space. A norm can be defined from the inner product, e.g.,
in H 1 (a, b), the norm is
Z b   1/2
2
kuk1 = u2 + u′ dx .
a

The distance in H m (Ω) (of general dimension) is defined as

d(u, v)m = ku − vkm . (7.26)

7.3.4 Relations between C m (Ω) and H m (Ω) — the Sobolev


embedding theorem
In 1D spaces, we have

H 1 (Ω) ⊂ C 0 (Ω), H 2 (Ω) ⊂ C 1 (Ω), · · · , H 1+j (Ω) ⊂ C j (Ω) .

Theorem 7.1. The Sobolev embedding theorem: If 2m > n, then

H m+j ⊂ C j , j = 0, 1, · · · , (7.27)

where n is the dimension of the independent variables of the elements in the Sobolev space.

Example 7.5. In 2D spaces, we have n = 2. The condition 2m > n means that m > 1.
From the embedding theorem, we have

H 2+j ⊂ C j , j = 0 =⇒ H 2 ⊂ C 0 , j = 1 =⇒ H 3 ⊂ C 1 , · · · . (7.28)

If u(x, y) ∈ H 2 , which means that u, ux , uy , uxx , uxy and uyy all belong to L2 , then
u(x, y) is continuous, but ux and uy may not be continuous!

Example 7.6. In 3D spaces, we have n = 3 and the condition 2m > n means that
m > 3/2 whose closest integer is number two, leading to the same result as in 2D:

H 2+j ⊂ C j , j = 0 =⇒ H 2 ⊂ C 0 , j = 1 =⇒ H 3 ⊂ C 1 , · · · . (7.29)

We regard the regularity of a solution as the degree of smoothness for a class of


problems measured in C m or H m space. Thus for u(x) ∈ H m or u(x) ∈ C m , the larger
the m the smoother the function.

i i

i i
164 Chapter 7. Theoretical Foundations of the Finite Element Method

7.4 FE analysis for 1D boundary value problems


For the simple 1D model problem

−u′′ = f , 0 < x < 1, u(0) = u(1) = 0 ,

we know that the weak form is


Z 1 Z 1
u′ v ′ dx = f v dx or (u′ , v ′ ) = (f, v) .
0 0

Intuitively, because v is arbitrary we can take v = f or v = u to get


Z 1 Z 1 Z 1 Z 1
′ ′ ′2
u v dx = u dx , f v dx = f 2 dx ,
0 0 0 0

so u, u′ , f , v and v ′ should belong to L2 (0, 1); i.e., we have u ∈ H01 (0, 1) and v ∈ H01 (0, 1);
so the solution is in the Sobolev space H01 (0, 1). We should also take v in the same space
for a conforming finite element method. From the Sobolev embedding theorem, we also
know that H 1 ⊂ C 0 , so the solution is continuous.

7.4.1 Conforming FE methods

Definition 7.2. If the finite element space is a subspace of the solution space, then the
finite element space is called a conforming finite element space, and the finite element
method is called a conforming FE method.

For example, the piecewise linear function over a given a mesh is a conforming
finite element space for the model problem. We mainly discuss conforming finite element
methods in this book. On including the boundary condition, we define the solution space
as

H01 (0, 1) = v(x) , v(0) = v(1) = 0 , v ∈ H 1 (0, 1) . (7.30)

When we look for a finite element solution in a finite dimensional space Vh , it should be a
subspace of H01 (0, 1) for a conforming finite element method. For example, given a mesh
for the 1D model, we can define a finite dimensional space using piecewise continuous
linear functions over the mesh:

Vh = {vh , vh (0) = vh (1) = 0 , vh is continuous piecewise linear} .

The finite element solution would be chosen from the finite dimensional space Vh , a
subspace of H01 (0, 1). If the solution of the weak form is in H01 (0, 1) but not in the Vh space;
then an error is introduced on replacing the solution space with the finite dimensional
space. Nevertheless, the finite element solution is the best approximation in Vh in some
norm, as discussed later.

i i

i i
7.4. FE analysis for 1D boundary value problems 165

7.4.2 FE analysis for 1D Sturm-Liouville problems


A 1D Sturm-Liouville problem on (xl , xr ) with a Dirichlet boundary condition at two ends
is
−(p(x)u′ (x))′ + q(x)u(x) = f (x) , xl < x < xr ,
u(xl ) = 0 , u(xr ) = 0 , (7.31)
p(x) ≥ pmin > 0 , q(x) ≥ qmin ≥ 0 .
The conditions on p(x) and q(x) guarantee the problem is well-posed, such that the
weak form has a unique solution. It is convenient to assume p(x) ∈ C(xl , xr ) and
q(x) ∈ C(xl , xr ). Later we will see that these conditions together with f (x) ∈ L2 (xl , xr ),
guarantee the unique solution to the weak form of the problem. To derive the weak form,
we multiply both sides of the equation by a test function v(x), v(xl ) = v(xr ) = 0 and
integrate from xl to xr to get
Z xr Z
xr

xr

−(p(x)u′ )′ + qu v dx = −pu′ v + pu′ v ′ + quv dx
xl xl xl
Z xr
= f v dx
xl
Z Z
xr
 xr
=⇒ pu′ v ′ + quv dx = f v dx , ∀v ∈ H01 (xl , xr ) or a(u, v) = L(v).
xl xl

7.4.3 The bilinear form


The integral
Z xr

a(u, v) = pu′ v ′ + quv dx (7.32)
xl

is a bilinear form, because it is linear for both u and v from the following
Z xr

a(αu + βw, v) = p(αu′ + βw′ )v ′ + q(αu + βw)v dx
xl
Z Z
xr
 xr

=α pu′ v ′ + quv dx + β w′ v ′ + qwv dx
xl xl

= αa(u, v) + βa(w, v) ,
where α and β are scalars; and similarly,
a(u, αv + βw) = αa(u, v) + βa(u, w) .
It is noted that this bilinear form is an inner product, usually different from the L2 and
H 1 inner products, but if p ≡ 1 and q ≡ 1 then
a(u, v) = (u, v).
Since a(u, v) is an inner product, under the conditions: p(x) ≥ pmin > 0, q(x) ≥ 0, we
can define the energy norm as
Z    12
p xr
′ 2 2
kuka = a(u, u) = p(u ) + qu dx , (7.33)
xl

i i

i i
166 Chapter 7. Theoretical Foundations of the Finite Element Method

where the first term may be interpreted as the kinetic energy and the second term as the
potential energy. The Cauchy-Schwartz inequality implies |a(u, v)| ≤ kuka kvka .
The bilinear form combined with linear form often simplifies the notation for the
weak and minimization forms, e.g., for the above Sturm-Liouville problem the weak form
becomes

a(u, v) = L(v) , ∀v ∈ H01 (xl , xr ) , (7.34)

and the minimization form is


n o
1
min F (v) = min a(v, v) − L(v) . (7.35)
v∈H01 (xl ,xr ) v∈H01 (xl ,xr ) 2

Later we will see that all self-adjoint differential equations have both weak and minimiza-
tion forms, and that the finite element method using the Ritz form is the same as with
the Galerkin form.

7.4.4 The FE method for 1D Sturm-Liouville problems using hat


basis functions
Consider any finite dimensional space Vh ⊂ H01 (xl , xr ) with the basis

φ1 (x) ∈ H01 (xl , xr ) , φ2 (x) ∈ H01 (xl , xr ) , · · · , φM (x) ∈ H01 (xl , xr ),

that is,

Vh = span {φ1 , φ2 , · · · , φM }
( )
X
M

= vs , vs = αi φi ⊂ H01 (xl , xr ) .
i=1

The Galerkin finite method assumes the approximate solution to be

X
M

us (x) = αj φj (x) , (7.36)


j=1

and the coefficients {αj } are chosen such that the weak form

a(us , vs ) = (f, vs ), ∀ vs ∈ Vh

is satisfied. Thus we enforce the weak form in the finite dimensional space Vh instead of
the solution space H01 (xl , xr ), which introduces some error.
Since any element in the space is a linear combination of the basis functions, we have

a (us , φi ) = (f, φi ) , i = 1, 2, · · · , M ,

or
!
X
M

a αj φj , φi = (f, φi ) , i = 1, 2, · · · , M ,
j=1

i i

i i
7.4. FE analysis for 1D boundary value problems 167

or

X
M

a(φj , φi ) αj = (f, φi ) , i = 1, 2, · · · , M .
j=1

In the matrix-vector form AX = F , this system of algebraic equations for the coefficients
is

    
a(φ1 , φ1 ) a(φ1 , φ2 ) ··· a(φ1 , φM ) α1 (f, φ1 )
    
 a(φ2 , φ1 ) a(φ2 , φ2 ) ··· a(φ2 , φM )     (f, φ2 ) 
   α2   
  = ,
 .. .. .. ..   ..   .. 
 . . . .  .   . 
    
a(φM , φ1 ) a(φM , φ2 ) ··· a(φM , φM ) αM (f, φM )

and the system has some attractive properties.

• The coefficient matrix A is symmetric, i.e., {aij } = {aji } or A = AT , since a(φi , φj ) =


a(φj , φi ). Note that this is only true for a self-adjoint problem such as the above,
with the second order ODE

−(pu′ )′ + qu = f .

For example, the similar problem involving the ODE

−u′′ + u′ = f

is not self-adjoint; and the Galerkin finite element method using the corresponding
weak form

(u′ , v ′ ) + (u′ , v) = (f, v) or (u′ , v ′ ) + (u, v ′ ) = (f, v)

produces terms such as (φ′i , φj ) that differ from (φ′j , φi ), so that the coefficient matrix
A is not symmetric.

• A is positive definite, i.e.,

xTA x > 0 if x 6= 0 , and all eigenvalues of A are positive.

i i

i i
168 Chapter 7. Theoretical Foundations of the Finite Element Method

Proof: For any η 6= 0, we show that η TA η > 0 as follows

X
M
X
M

η TA η = η T (Aη) = ηi aij ηj
i=1 j=1

X
M
X
M

= ηi a(φi , φj )ηj
i=1 j=1

X
M X
M

= ηi a(φi , ηj φj )
i=1 j=1
!
X
M
X
M

= ηi a φi , ηj φj
i=1 j=1
!
X
M
X
M

=a ηi φi , ηj φj
i=1 j=1

= a (vs , vs ) = kvs k2a > 0 ,

X
M

since vs = ηi φi 6= 0 because η is a nonzero vector and the {φi }’s are linear independent.
i=1

7.4.5 Local stiffness matrix and load vector using the hat basis
functions
The local stiffness matrix using the hat basis functions is a 2 × 2 matrix of the following,
" R xi+1 R xi+1 # " R xi+1 R xi+1 #
xi
p(x) (φ′i )2 dx xi
p(x) φ′i φ′i+1 dx xi
q(x) φ2i dx xi
q(x) φi φi+1 dx
Kie = R xi+1 R xi+1 + R xi+1 R xi+1 ,
xi
p(x) φ′i+1 φ′i dx xi
p(x) (φ′i+1 )2 dx xi
q(x) φi+1 φi dx xi
q(x) φ2i+1 dx

and the local load vector is


" R xi+1 #
xi
f φi dx
Fie = R xi+1 .
xi
f φi+1 dx

The global stiffness matrix and load vector can be assembled element by element.

7.5 Error analysis of the FE method


Error analysis for finite element methods usually includes two parts:
1. error estimates for an intermediate function in Vh , often the interpolation function;
and
2. convergence analysis, a limiting process that shows the finite element solution con-
verges to the true solution of the weak form in some norm, as the mesh size h
approaches zero.
We first recall some notations and setting up:

i i

i i
7.5. Error analysis of the FE method 169

1. Given a weak form a(u, v) = L(v) and a space V , which usually has infinite dimen-
sion, the problem is to find a u ∈ V such that the weak form is satisfied for any
v ∈ V . Then u is called the solution of the weak form.

2. A finite dimensional subspace of V denoted by Vh (i.e. Vh ⊂ V ) is adopted for a


conforming finite element method and it does not have to depend on h, however.

3. The solution of the weak form in the subspace Vh is denoted by uh , i.e., we require
a(uh , vh ) = L(vh ) for any vh ∈ Vh .

4. The global error is defined by eh = u(x) − uh (x), and we seek a sharp upper bound
for keh k using certain norm.

It was noted that error is introduced when the finite dimensional space replaces the solution
space, as the weak form is usually only satisfied in the sub-space Vh and not in the solution
space V . However, we can prove that the solution satisfying the weak form in the sub-
space Vh is the best approximation to the exact solution u in the finite dimensional space
in the energy norm.

u u − uh
u − uh
u
Vh
uh
Vh
uh

Figure 7.2. A diagram of FE approximation properties. The finite element


solution is the best approximation to the solution u in the finite dimensional space Vh
in the energy norm; and the error u − uh is perpendicular to the finite dimensional
space Vh in the inner product of a(u, v).

Theorem 7.3.

1. uh is the projection of u onto Vh through the inner product a(u, v), i.e.,

u − uh ⊥ Vh or u − uh ⊥ φi , i = 1, 2, · · · , M , (7.37)

a(u − uh , vh ) = 0 ∀ vh ∈ Vh or a(u − uh , φi ) = 0 , i = 1, 2, · · · , M , (7.38)

where {φi }’s are the basis functions.

2. uh is the best approximation in the energy norm, i.e.,

ku − uh ka ≤ ku − vh ka , ∀ vh ∈ Vh .

i i

i i
170 Chapter 7. Theoretical Foundations of the Finite Element Method

Proof:

a(u, v) = (f, v) , ∀ v ∈ V,
→ a(u, vh ) = (f, vh ) , ∀ vh ∈ Vh since Vh ⊂ V ,
a(uh , vh ) = (f, vh ) , ∀ vh ∈ Vh since uh is the solution in Vh ,
subtract → a(u − uh , vh ) = 0 or a(eh , vh ) = 0 , ∀ vh ∈ Vh .

Now we prove that uh is the best approximation in Vh .

ku − vh k2a = a(u − vh , u − vh )
= a(u − uh + uh − vh , u − uh + uh − vh )
= a(u − uh + wh , u − uh + wh ) , on letting wh = uh − vh ∈ Vh ,
= a(u − uh , u − uh + wh ) + a(wh , u − uh + wh )
= a(u − uh , u − uh ) + a(u − uh , wh ) + a(wh , u − uh ) + a(wh , wh )
= ku − uh k2a + 0 + 0 + kwh k2a , since, a(eh , uh ) = 0
≥ ku − uh k2a

i.e., ku − uh ka ≤ ku − vh ka . Fig. 7.2 is a diagram to illustrate the theorem.

Example: For the Sturm-Liouville problem,


Z b 
ku − uh k2a = p(x) (u′ − u′h )2 + q(x) (u − uh )2 dx
a
Z b Z b
≤ pmax (u′ − u′h )2 dx + qmax (u − uh )2 dx
a a
Z b 
≤ max{pmax , qmax } (u′ − u′h )2 + (u − uh )2 dx
a

= Cku − uh k21 ,

where C = max{pmax , qmax }. Thus we obtain

ku − uh ka ≤ Ĉku − uh k1 ,
ku − uh ka ≤ ku − vh ka ≤ C̄ku − vh k1 .

7.5.1 Interpolation functions and error estimates


Usually the solution is unknown; so in order to get the error estimate we choose a special
vh∗ ∈ Vh , for which we can get a good error estimate. We may then use the error estimate
ku − uh ka ≤ ku − vh∗ ka to get an error estimate for the finite element solution (maybe over-
estimated). Usually we can choose a piecewise interpolation function for this purpose.
That is another reason that we choose piecewise linear, quadratic or cubic functions over
the given mesh in finite element methods.

i i

i i
7.5. Error analysis of the FE method 171

Linear 1D piecewise interpolation function


Given a mesh x0 , x1 , x2 , · · · , xM , the linear 1D piecewise interpolation function is defined
as
x − xi x − xi−1
uI (x) = u(xi−1 ) + u(xi ) , xi−1 ≤ x ≤ xi .
xi−1 − xi xi − xi−1

It is obvious that uI (x) ∈ Vh , where Vh ⊂ H 1 is the set of continuous piecewise linear


functions that have the first order weak derivative, so

ku − uh ka ≤ ku − uI ka .

Since u(x) is unknown, then so is uI (x). Nevertheless, we know the upper error bound of
the interpolation functions.

Theorem 7.4. Given a function u(x) ∈ C 2 [a, b] and a mesh x0 , x1 , x2 , · · · , xM , the


continuous piecewise linear function uI has the error estimates

h2 ′′
ku − uI k∞ = max |u(x) − uI (x)| ≤ ku k∞ , (7.39)
x∈[a,b] 8

ku′ (x) − u′I (x)kL2 (a,b) ≤ h b − aku′′ k∞ . (7.40)

Proof: If e˜h = u(x) − uI (x), then e˜h (xi−1 ) = e˜h (xi ) = 0. From Rolle’s theorem,
there must be at least one point zi between xi−1 and xi such that e˜h ′ (zi ) = 0, hence
Z x
e˜h ′ (x) = e˜h ′′ (t) dt
zi
Z x

= u′′ (t) − u′′I (t) dt
zi
Z x
= u′′ (t) dt .
zi

Therefore, we obtain the error estimates below


Z x Z x
|e˜h ′ (x)| ≤ |u′′ (t)|dt ≤ ku′′ k∞ dt ≤ ku′′ k∞ h , and
zi zi
 Z b
 21
′ ′

ke˜h kL2 (a,b) = ke˜h k0 ≤ ku′′ k2∞ 2
h dt ≤ b − aku′′ k∞ h ;
a

so we have proved the second inequality. To prove the first, assume that xi−1 + h/2 ≤
zi ≤ xi , otherwise we can use the other half interval. From the Taylor expansion

e˜h (x) = e˜h (zi + x − zi ) , assuming xi−1 ≤ x ≤ xi ,


1
= e˜h (zi ) + e˜h ′ (zi )(x − zi ) + e˜h ′′ (ξ)(x − zi )2 , xi−1 ≤ ξ ≤ xi ,
2
1 ′′ 2
= e˜h (zi ) + e˜h (ξ)(x − zi ) ,
2

i i

i i
172 Chapter 7. Theoretical Foundations of the Finite Element Method

so at x = xi we have
1 ′′
0 = e˜h (xi ) = e˜h (zi ) + e˜h (ξ)(xi − zi )2 ,
2
1
e˜h (zi ) = − e˜h ′′ (ξ)(xi − zi )2 ,
2
1 ′′ h2 ′′
|e˜h (zi )| ≤ ku k∞ (xi − zi )2 ≤ ku k∞ .
2 8
Note that the largest value of e˜h (x) has to be the zi where the derivative is zero.

7.5.2 Error estimates of the finite element methods using the


interpolation function

Theorem 7.5. For the 1D Sturm-Liouville problem, the following error estimates hold,

ku − uh ka ≤ Chku′′ k∞ ,

ku − uh k1 ≤ Ĉhku′′ k∞ ,

where C and Ĉ are two constants.

Proof:

ku − uh k2a ≤ ku − uI k2a
Z b 
≤ p(x) (u′ − u′I )2 + q(x) (u − uI )2 dx
a
Z b 
≤ max {pmax , qmax } (u′ − u′I )2 + (u − uI )2 dx
a
Z b

≤ max {pmax , qmax } ku′′ k2∞ h2 + h4 /64 dx
a

≤ Ch2 ku′′ k∞ .

The second inequality is obtained because k ka and k k1 are equivalent, so

ckvka ≤ kvk1 ≤ Ckvka , ĉkvk1 ≤ kvka ≤ Ĉkvk1 .

7.5.3 Error estimate in the pointwise norm


We can easily prove the following error estimate.

Theorem 7.6. For the 1D Sturm-Liouville problem,

ku − uh k∞ ≤ Chku′′ k∞ , (7.41)

where C is a constant. The estimate is not sharp (or optimal); or simply it is over-
estimated.

i i

i i
7.6. Exercises 173

We note that u′h is discontinuous at nodal points, and the infinity norm ku′ − u′h k∞ can
only be defined for continuous functions.
Proof:
Z x
eh (x) = u(x) − uh (x) = e′h (t)dt
a
Z b
|eh (x)| ≤ |e′h (t)|dt
a
Z b 1/2 Z b 1/2
≤ |e′h |2 dt 1 dt
a a
Z b 1/2
√ p
≤ b−a |e′h |2 dt
a
pmin
r
b−a
≤ keh ka
pmin
r
b−a
≤ ke˜h ka
pmin
≤ Ch ku′′ k∞ .

Remark 7.1. Actually, we can prove a better inequality

ku − uh k∞ ≤ Ch2 ku′′ k∞ ,

so the finite element method is second order accurate. This is an optimal (sharp) error
estimate.

7.6 Exercises
1. Assuming the number of variables n = 3, describe the Sobolev space H 3 (Ω) (i.e.,
for m = 3) in terms of L2 (Ω), retaining all the terms but not using the multi-
index notation. Then using the multi-index notation when applicable, represent the
inner product, the norm, the Schwartz inequality, the distance, and the Sobolev
embedding theorem in this space.

2. Consider the function v(x) = |x|α on Ω = (−1, 1) with α ∈ R. For what values
of α is v ∈ H 0 (Ω)? (Consider both positive and negative α.) For what values is
v ∈ H 1 (Ω)? in H m (Ω)? For what values of α is v ∈ C m (Ω)?
Hint: Make use of the following
(
α
xα if x ≥ 0
|x| =
α
(−x) if x < 0 ,

and when k is a non-negative integer note that


(
α
x2k if α = 2k
|x| =
1 if α = 0 ;

i i

i i
174 Chapter 7. Theoretical Foundations of the Finite Element Method

also
 
 0 if α > 0 Z 2
 1  if α > −1
α
lim |x| = 1 if α = 0 and α
|x| dx = α+1
x→0 
 
−1 ∞ if α ≤ −1 .
∞ if α < 0

3. Are each of the following statements true or false? Justify your answers.
(a) If u ∈ H 2 (0, 1), then u′ and u′′ are both continuous functions.
(b) If u(x, y) ∈ H 2 (Ω), then u(x, y) may not have continuous partial derivatives
∂u/∂x and ∂u/∂y.
Does u(x, y) have first and second order weak derivatives?
Is u(x, y) continuous in Ω?

4. Consider the Sturm-Liouville problem


′
− p(x)u(x)′ + q(x)u(x) = f (x) , 0 < x < π,

αu(0) + βu′ (0) = γ , u′ (π) = ub ,


where 0 < pmin ≤ p(x) ≤ pmax < ∞ ,
0 ≤ qmin ≤ q(x) ≤ qmax < q∞ .

(a) Derive the weak form for the problem. Define a bilinear form a(u, v) and a
linear form L(v) to simplify the weak form. What is the energy norm?
(b) What kind of restrictions should we have for α, β, and γ in order that the
weak form has a solution?
(c) Determine the space where the solution resides under the weak form.
(d) If we look for a finite element solution in a finite dimensional space Vh using
a conforming finite element method, should Vh be a subspace of C 0 , or C 1 , or
C2?
(e) Given a mesh x0 = 0 < x1 < x2 · · · < xM −1 < xM = π, if the finite dimensional
space is generated by the hat functions, what kind of structure do the local
and global stiffness matrix and the load vector have? Is the resulting linear
system of equations formed by the global stiffness matrix and the load vector
symmetric, positive definite and banded?
5. Consider the two-point BVP

−u′′ (x) = f (x), a<x<b, u(a) = u(b) = 0 .

Let uh (x) be the finite element solution using the piecewise linear space (in H01 (a, b))
spanned by a mesh {xi }. Show that

ku − uh k∞ ≤ Ch2 ,

where C is a constant.
Hint: First show kuh − uI ka = 0, where uI (x) is the interpolation function in Vh .

i i

i i